What you will learn in Effective Conversational AI
Chatbots that work
Chapter by chapter
Chapter 9: Harnessing context for an adaptive virtual assistant experience
What you learn in chapter 9 →Concepts explained in this book
- language modelA large language model is a machine learning model trained on very large text datasets to model and generate sequences of language.
- retrieval-augmented generationRetrieval-augmented generation is a technique that conditions generated responses on information retrieved from external documents or data sources.
- intentionA user intent is the goal, purpose, or desired action that a user expresses or implies in a conversational interaction.
- evaluation datasetA test set, or test data, is a collection of evaluation inputs used to measure a model or system's behavior on cases separate from its training examples.
- classifierClassification is the process of assigning an input to one or more predefined categories based on its features or content.
- generative modelingAn AI model that creates new content, such as text, images, audio, or code, from learned patterns or conditioning information.
- promptA prompt is an input supplied to a language model that provides a task, instructions, context, or other conditioning information.
- contextContext is information about a user or interaction that a conversational system retains to interpret and manage the conversation.
- conversational aiA virtual assistant is software that conducts natural-language interactions to answer questions or perform tasks for a user.
- accuracyAn accuracy rate is the proportion of evaluated predictions or answers that are correct.
- handoffEscalation is the transfer of a chatbot interaction to a human agent or other additional handling when the bot cannot complete the request or the user requests assistance.
- training datasetTraining data, or a training dataset, is a collection of examples used to fit a machine-learning model.
- personalizationPersonalization is adapting a conversational response to an individual user's circumstances, preferences, or context.
- conversational flowA conversational flow is the structured sequence of dialogue turns and information exchanges in an interaction.
- opt-out requestAn opt-out is a user's decision to end or abandon an interaction with an automated conversational system instead of continuing it.
- workflowA task flow is an ordered sequence of conversational steps that accomplishes a specific user task.
- speech recognitionSpeech-to-text, or automatic speech recognition, converts spoken audio into written text.
- question answeringQuestion-answering is a conversational AI capability that responds directly to a user's question, usually without follow-up interaction.
- hallucinationAn AI hallucination is generated content that presents fabricated, unsupported, or incorrect information as if it were reliable.
- recallRecall is the proportion of relevant or correct cases that a classifier successfully identifies.
- utteranceAn utterance is a piece of natural-language input produced by a user in a conversational interaction.
- vector databaseA vector store is a storage system that indexes numerical vector representations for similarity search.
- data modalityThe communication mode through which a user interacts with a conversational system, such as text, voice, email, or push notification.
- precisionPrecision is the proportion of a classifier's positive or selected predictions that are correct.
- dialogueChat history is the recorded sequence of messages exchanged during a user's conversational interactions.
- few-shot learningFew-shot prompting is the practice of providing a language model with a small number of example demonstrations in a prompt to guide its response.
- optimizationContinuous improvement is an iterative cycle of evaluating a system, identifying shortcomings, applying changes, and reassessing results.
- contractAn application programming interface (API) is a defined interface through which one software system or component requests capabilities, data, or operations from another.
- resultA conversation outcome is the classified result of a chatbot interaction, such as successful resolution, failure, or rejection of chatbot assistance.
- completion rateTask completion rate, or success rate, is the proportion of user tasks or interactions completed successfully.
- conversation abandonmentConversation abandonment is the ending of a conversational interaction before the user's intended task or outcome is completed.
- botA chatbot is a software agent that conducts conversations with users through text or other communication channels.
- disambiguationDisambiguation is a conversational technique that presents alternative interpreted goals so a user can select or clarify the intended one.
- context-awarenessSemantic understanding is the computational interpretation of the meaning conveyed by language.
- information retrievalThe retrieval phase is the stage of a retrieval-augmented generation process that finds and obtains information relevant to an input.
- process-oriented botA process-oriented bot is a conversational agent that guides a user through a sequence of questions or actions to achieve a goal or complete a transaction.
- one-shot promptingOne-shot prompting is the practice of providing a language model with one example demonstration in a prompt to guide its response.
- intent-based chatbotAn intent-based chatbot is a conversational bot that maps user queries to predefined intent categories and follows associated responses or dialogue paths.
- summarizationA summary is a shortened representation of content that preserves its key highlights and details.
- self-serviceA service interaction in which customers complete a task or obtain information without assistance from a human agent.
- samplingSampling is a language-generation decoding strategy that selects output tokens probabilistically from a model's predicted distribution.
- guardrailA guardrail is a constraint, control, or oversight mechanism that limits unsafe, unsuitable, or undesired behavior by an AI system.
- complexityThe degree of structural difficulty in a conversational task or interaction, determined by the number, interdependence, length, branching, or sequencing of steps.
- prompt engineeringPrompt engineering is the design of instructions, context, examples, and input data supplied to a language model to influence its output.
- old aiTraditional AI is an approach to artificial intelligence that relies primarily on predefined rules, classifications, or explicitly programmed logic rather than open-ended content generation.
- relevanceAnswer relevance is the degree to which an answer directly addresses the user's question and context.
- one-shot exampleA one-shot example is a single input-output demonstration supplied to a model to guide its behavior on a task.
- detailed outcome modelA detailed outcome model is a hierarchical categorization of conversational results, reasons, and higher-level success or failure categories.
- usability metricEffectiveness is the degree to which a chatbot achieves its intended goals while providing users with a positive and efficient experience.
- text chunkA coherent segment of text treated as an individual unit for embedding, indexing, or retrieval.
- response generationAnswer generation is the production of a candidate response from a language model based on retrieved information and instructions.
- embeddingA text embedding is a numerical vector representation that encodes semantic or contextual properties of text.
- temperatureTemperature is a generation parameter that controls the degree of randomness in a language model's token selection.
- proactive supportAssistance initiated by a service or conversational system to address likely user needs or potential problems before the user explicitly asks for help.
- cross-validationK-fold cross-validation is a model-evaluation technique that partitions a dataset into k folds and repeatedly evaluates a model with each fold held out in turn.
- retryA retry is a conversational handling path that asks the user to provide input again after an attempted response is not accepted.
- fallbackA fallback path is a conversational handling route for input that cannot be understood or handled by the primary flow.
- adaptive flowAn adaptive flow is a conversational interaction sequence that dynamically responds to a user's context, preferences, history, or changing circumstances.
- fidelityFaithfulness is a generation metric measuring whether an answer is factually supported by the retrieved context.
- baselineA baseline is a reference measurement of system performance taken before a change is made.
- throughputThroughput is the amount of work or number of requests a system processes per unit of time.
- ambiguous intentAn ambiguous utterance is conversational input that is understandable but does not provide enough information to determine a single user goal.
- golden intentA golden intent is the authoritative intent label assigned to a conversational input for evaluation or testing.
- multi-intent utteranceA multi-intent utterance is conversational input that expresses two or more distinct user goals.
- negative testingNegative testing evaluates whether a system correctly rejects, declines, or otherwise handles inputs that it should not answer or process.
- out-of-scope utteranceAn out-of-scope utterance expresses a clear goal relevant to a domain but not supported by the current conversational solution.
- unintelligible utteranceAn unintelligible utterance is conversational input whose meaning cannot be reliably determined because it is incoherent, severely corrupted, or otherwise incomprehensible.
- retrieval metricRetrieval accuracy is the degree to which a retrieval system fetches the most relevant information for a query.
- containmentContainment is the proportion or number of customer interactions completed without human involvement.
- f-scoreAn F1 score is a classification metric that combines precision and recall through their harmonic mean.
- continuous improvementA feedback loop is a process in which observed results or user information are fed back into a system to refine its behavior.
- greeting messageA conversation entry point is the initial interaction through which a user enters a conversational system and receives orientation about what it can do.
- cost-benefit analysisCost-benefit analysis is a comparison of the expected advantages of a change with the resources and expenses required to implement it.
- data annotationData annotation is the process of assigning labels or judgments to data examples for training, testing, or evaluation.
- user surveyA survey is a structured instrument for collecting responses about people's experiences, opinions, or needs.
- chatbot platformA software platform that provides capabilities for building conversational applications, including natural language understanding and dialogue management.
- session historySession history is the contents of the current conversational session, retained for immediate context and typically discarded after the session ends.
- zero-shot promptingA zero-shot prompt is a prompt that specifies a task without providing an example of the desired input-output behavior.
- business goalA business objective is a desired organizational result that a conversational system is intended to achieve.
- FAISSFAISS is a software library for efficient similarity search and clustering of dense vectors.
- intent classificationIntent classification is the assignment of a user utterance to one of a set of predefined intent categories.
- customer problemA user pain point is a specific difficulty, frustration, or unmet need experienced by a user during an interaction with a system or service.
- speech mistranscriptionA speech mistranscription is an incorrectly recognized word or phrase produced when spoken language is converted into text.
- greedy algorithmGreedy decoding is a language-generation strategy that selects the highest-scoring next token at each step.
- voice interfaceVoice modality is an interaction mode in which users communicate with a system through spoken language and receive auditory responses.
- output formatStructured output is generated information organized according to a predefined machine-readable format or set of fields.
- user experienceUser experience is the overall quality of a person's interaction with a product, service, or system, including how natural and easy the interaction feels.
- context windowA context window is the maximum amount of input text or tokens that a language model can process in one inference request.
- document loaderA document loader is a component that reads content from a configured source or file format and represents it as documents with text and metadata.
- document structureDocument structure is the organization of content into elements such as headings, tables, lists, and their relationships.
- document transformerA document transformer is a component that preprocesses or restructures documents for downstream indexing, retrieval, or language-model processing.
- LangChainLangChain is a software library for developing applications that compose language models with prompts, data sources, retrieval, and other processing components.
- text splitterA text splitter is a component or method that divides text into chunks according to structural, character, code, markup, or token-based rules.
- knowledge baseA knowledge base is an organized collection of information that can be searched to support answering questions.
- error handlingError handling is the management of recognition, interpretation, or execution failures so that a system can recover or guide the user toward an appropriate outcome.
- conversational outcome metricA conversational outcome metric is a measurement of what happens during or after a conversational interaction, such as completion, abandonment, or escalation.
- effect-effort matrixAn effect-effort matrix is a prioritization model that categorizes improvement opportunities by their expected effect and implementation effort.
- solutioningSolutioning is the process of outlining a high-level fix by defining its responsibilities, required changes, implementation approach, and effort.
- task prioritizationProblem prioritization is the process of ranking problems according to their expected value, effect, frequency, cost, and effort to resolve.
- triageProblem triage is the systematic evaluation and prioritization of identified system problems according to factors such as effect, frequency, and resolution effort.
- hybrid approachA hybrid approach combines generative AI capabilities with human-builder control in the creation or improvement of conversational AI systems.
- representative blind test setA representative blind test set is an evaluation dataset drawn from real user inputs and annotated with the correct labels without being used during model development, while preserving the frequency distribution of topics in production.
- human-in-the-loopHuman review is the examination of system outputs by people who assess their accuracy and contextual appropriateness.
- context variableA context variable is a named value that carries information relevant to processing a conversational interaction.
- context-aware chatbotA context-aware chatbot is a chatbot that retains and applies relevant information about the user and interaction to interpret requests and tailor assistance.
- persistent user historyPersistent user history is a user's interaction history retained across multiple sessions and potentially across communication modalities.
- benchmarkBenchmarking is the systematic measurement and comparison of a system against defined performance criteria or alternatives.
- document groundingDocument grounding is the practice of basing generated responses on identifiable source documents or passages relevant to the query.
- ElasticsearchElasticsearch is a distributed search and analytics engine that indexes and retrieves structured and unstructured data.
- evaluation pipelineAn evaluation pipeline is an organized sequence of data preparation, system execution, measurement, and analysis steps for assessing a model or application.
- MTEBMTEB is a benchmark suite and leaderboard for evaluating text embedding models across multiple tasks.
- prompt injectionPrompt injection is an attempt to manipulate a language model's instructions or behavior through specially crafted input.
- user-centric designUser design is an approach that centers an interactive system on users’ goals, needs, constraints, and likely expressions.
- conversational searchConversational search is an interaction pattern that retrieves information through natural-language dialogue and presents it as a conversational response.
- meeting minutesA conversation summary is a concise representation of a complete conversational interaction that combines essential structured information with a short text account.
- structured summaryA structured summary is a condensed representation of source information organized into predefined fields or an explicit data structure.
- conversation designConversational design is the practice of structuring dialogue, interaction steps, and responses to support user goals.
- JSONJSON is a text-based data-interchange format that represents structured values using objects, arrays, and primitive values.
- sentence fragmentA sentence fragment is an incomplete grammatical sequence that conveys part of a sentence's meaning without forming a complete sentence.
- small changeA narrowly scoped system modification whose effects and causes can be evaluated with relatively low risk.
- AI-generated dataA generated test dataset is a collection of evaluation examples produced artificially, often by prompting a generative model to create variations of existing examples.
- blind testingBlind testing evaluates a classifier on utterances that were not included in the data used to train it.
- conversational actionA conversational action is a configured operation that starts under specified conversational conditions and performs a response or subsequent interaction step.
- intent-based question-answering systemAn intent-based question-answering system maps recognized intent categories to builder-defined responses.
- question-answering actionA question-answering action is a conversational action that detects an intent and provides an answer.
- static question-answering actionA static question-answering action detects an intent and returns a fixed response without conditional logic or a subsequent step.
- feedbackUser feedback is information provided by users about their experiences, needs, or evaluations of a software system.
- multimodal interactionMultimodal support is the provision of conversational assistance through two or more interaction modalities, such as voice, text, or visual content.
- semantic similaritySemantic similarity is the degree of shared meaning between two pieces of text, typically estimated from their representations.
- hybrid chatbotA hybrid model combines intent-based routing with search or generative-AI capabilities in one conversational system.
- confusion matrixA confusion matrix is a tabular summary comparing a classifier's predicted categories with the authoritative categories for evaluation examples.
- scalabilityScalability is the ability of a system to maintain acceptable operation as data volume, query volume, or workload increases.
- false positiveA false positive is a classification error in which a model predicts a category or condition that is incorrect for the input.
- ideal responseAn ideal response is a reference answer judged appropriate, factual, complete, and consistent with the intended conversational experience.
- metricA metric is a quantitative measure used to evaluate a system's performance or outcomes.
- entityAn entity is a contextual element in a user utterance that is relevant to interpreting an intent.
- preprocessingData preparation is the process of transforming collected data into a suitable form for testing or model training.
- free-text summaryA free-text summary is a concise natural-language account of the essential events and outcome of a conversation.
- screen popA screen pop is a contact-center software feature that displays contextual information to an agent while the agent handles a conversation.
- conversionA conversion rate is the proportion of eligible users or interactions that produce a desired business action.
- return on investmentReturn on investment is the financial return produced by an investment relative to its cost.
- revenue generationRevenue generation is the creation of income through sales, services, or other commercial activity.
- understandingChatbot understanding is a system's ability to correctly interpret user inputs and determine what the user wants.
- privacyPrivacy is the protection and appropriate handling of information about identifiable individuals.
- grammatical variationA grammatical variation is a change in the grammatical form of an utterance while preserving its underlying meaning.
- Chroma DBChroma is an open-source vector database and retrieval system for storing and searching embedding vectors.
- LuceneApache Lucene is an open-source search library that provides text indexing and information-retrieval capabilities.
- adaptabilityAdaptiveness is the ability of a conversational system to adjust its response to the user's wording, tone, context, or requested level of detail.
- comparison checkA comparison check is an evaluation of a generated answer against retrieved passages to identify unsupported content.
- grounded answerA grounded response is an answer whose claims are supported by retrieved source evidence.
- passage retrievalPassage retrieval is the process of finding text passages relevant to a query from a document collection.
- extractive summarizationExtractive summarization is a summarization technique that constructs a shorter representation by selecting or extracting information from source content.
- Mistral 7B Instructmistral-7b-instruct-v0-2 is an instruction-tuned large language model in the Mistral model family.
- decoder-only transformerA decoder-only architecture is a language-model architecture trained to predict successive tokens and generate text from preceding context.
- encoder-only transformerAn encoder-only architecture is a language-model architecture that transforms input text into contextual representations without generating output sequences.
- hybrid approachA hybrid conversational approach combines classification, predefined or task-oriented dialogue flows, and generated responses in one conversational system.
- encoder-decoder architectureAn encoder-decoder architecture is a language-model architecture that encodes an input sequence and decodes a separate output sequence.
- context managementThe ongoing process of tracking, retaining, updating, and applying conversational context as an interaction progresses.
- special tokenA special token is a reserved token with a control or structural meaning in a language-model input or output.
- instrumentationInstrumentation is the deliberate collection and storage of operational data points from a software interaction for later analysis or processing.
- webhookA webhook is an API endpoint that is invoked automatically when a specified event occurs in another system.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.